Alex Wiltschko - Giving Computers A Sense Of Smell - [Invest Like the Best, EP.415]
What this covers
In this conversation, Patrick tours Osmo's laboratory with founder Alex Wiltschko, who explains how the company teaches computers to read and write scent. Alex begins with why smell resisted digitization: no one has a periodic table of primary smells, Osmo's scent map has roughly 300 dimensions against color's three, and Moravec's Paradox predicts that evolution-honed senses lag behind chess and theorem proving. His answer combines Graph Neural Network methods borrowed from drug discovery with a read-write cycle in which every experiment yields new data. Osmo's first "teleported" scent, a fresh summer plum, serves as the demonstration.
The middle of the episode covers evidence and product. Alex describes a double-blind test with Monell in which Osmo's predictions on 400 novel molecules beat the average trained human panelist, which he treats as passing an Odor Turing Test. He then turns to Generation, Osmo's fragrance house, contrasting the usual 12-to-18-month fragrance brief cycle with a conversational interface that returns custom scents in minutes. The business logic gets attention too: fragrance sits in about 90 percent of household products, repurchase rates run high, and Osmo chose to compete directly rather than sell software to a few unready fragrance houses.
The later turns are more speculative. A miniaturized sniffer already distinguishes authentic from counterfeit sneakers within 20 seconds at StockX. Alex also calls health monitoring through breath and sweat the long-term goal, a claim he presents as a forecast and not a result. He closes by placing scent AI at the start of its technology S-curve, while text AI is flattening.
Osmo argues that digitizing smell is now feasible because the bottleneck was never the hardware but the missing software and proprietary data pipeline, and that scent design for fragrance brands is the commercial wedge that funds the longer path to health and security applications.
- Existing GC-MS hardware is already strong; the missing piece was the software mapping signals to molecules and scent, which Osmo built.
- Data is the fuel for AI on scent, so Osmo built its own labeling, sensing, and synthesis infrastructure because licensed data was not AI-compatible.
- Generation, a custom fragrance business, is the near-term revenue path, while health monitoring and counterfeit detection are later frontiers.
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
Being able to both read a smell into data and write it back out creates a virtuous cycle in which every run produces new data, and if the next day's experiments are chosen to teach the system more, this is active learning that makes AI systems improve quickly.
“If you can read and write, then you can create this virtuous cycle where you're creating data at every run of the loop.”
Smell is the most ancient sense, evolved from microbes sipping chemicals to find food, and it projects to memory and emotion centers such as the hippocampus and amygdala faster than any other sense.
“The number of steps it takes for that information to get to your centers of memory and your centers of emotion is faster than any other sense.”
In the traditional fragrance process, a buyer who rejects library samples and demands a custom scent faces a 12- to 18-month back-and-forth, with each round of notes taking about three months, and then must separately run application testing in the product.
“Now you're looking at a 12- to 18-month process of going back and forth. Every time you submit notes, it might take three months for them to get back to you.”
Graph neural networks fit molecules because molecules have arbitrary shapes, atom counts, and bond arrangements, unlike images or text which suit grids or sequences, making them the key technique linking AI and chemistry.
“molecules can have any shape, there can be any number of atoms, and the bonds can be all rearranged.”
Predicting what a molecule will smell like from its drawn structure, such as apple, cinnamon, or anise, has been a hundred-year-old unsolved problem in olfaction.
“It turns out that's a hundred-year-old problem that nobody has been able to solve.”
Moravec's Paradox holds that tasks easy for humans are hard for computers, because evolution optimized them, so smelling lagged while proving theorems and beating chess fell first; smell is now being cracked as robots learn to walk.
“the easier it is for a person to do, the harder it is for a computer.”
Technology follows an S-curve from poor performance through rapid improvement to plateau, text AI has reached the flattening phase after passing the Turing test, while scent AI is only just taking off at the far left of its S-curve.
“Technology usually proceeds on an S curve. It sucks. It's getting better. Oh, my gosh, it's getting better super fast. And then it levels off.”
Fragrance businesses are recession resistant because people trade down from luxury perfume to hand soap, fragrance appears in about 90 percent of household products, and customer repurchase rates exceed 50 percent once a supplier is won.
“Typically, if you've won the business, industry standards for repurchasing are well above 50%.”
The chemistry of blood and organs is exhaled through breath and sweat, so scent may carry health signals, which dogs already detect, and Osmo sees human health and wellness as its long-term holy grail.
“What's on the inside of us gets to the outside, what's in our blood and in our organs is eventually exhaled through our breath, our sweat, whatever.”
No one knows the equivalent of a periodic table or set of primary smells that can be combined to build any scent, and this absence has been the core reason scent was not digitized.
“So nobody knows. But we're teaching the machines to figure that out. That has been the core issue of why scent hasn't been digitized”
The binding constraint for AI progress in a new sense like smell is data rather than model architecture, so the real moat is the infrastructure that generates labeled physical-digital data.
“And the thing that is the fuel here is data.”
In a double-blind test with Monell, Osmo predicted the smells of hundreds of thousands of molecules, selected 400 structurally novel ones, kept predictions secret, and found its AI predictions beat the average trained human panelist.
“It turned out that our AI predictions of what these smells were going to be were better than the average panelist.”
Knowing that a molecule exists is not enough; creating a maple syrup variant with more cherry or avoiding an unsafe molecule requires tacit perfumer tradecraft, which Osmo aims to automate.
“That requires tons of tradecraft. That is what we're automating.”
Whether to sell platform software or to compete as an end-product company depends on whether the target market is ready to buy software; because fragrance houses are few and not ready, Osmo chose to enter the market itself, as with Metropolis in parking.
“The question is, do you become a software provider for the incumbents or do you take your capabilities and compete in that market?”
Licensed scent datasets were not AI-compatible, so Osmo had to build its own data pipeline from the ground up, including a staffed sensory labeling operation and sensor labs running around the clock.
“We tried to do some licensing deals. We were successful. The data wasn't AI compatible. So we said, 'Okay, I guess we have to create it ourselves.'”
A miniaturized sniffer trained on a few real and fake pairs of a sneaker SKU can distinguish authentic from counterfeit within 20 seconds, and it is deployed at StockX, with counterfeiters' visual replicas failing to match the shoe's olfactory fingerprint.
“It will, within 20 seconds, tell you whether it's real or fake.”
Because the AI predicted smells better than the average human panelist, adding a panelist would be less valuable than querying the software, so the model passed an odor Turing test without needing a physical molecule.
“you'd prefer to ask our software what it smells like, even without access to the physical molecule, than to train up another person to physically smell it—which is passing an odor Turing test.”
The traditional fragrance brief process, which can take 12 to 18 months with multiple rounds and application testing, will be replaced by a conversational AI interface that produces custom scents in minutes, with Generation as Osmo's fragrance house.
“this is a multi-month process that we're condensing down into minutes.”
Getting good at designing fruit, flower, and vegetable scents transfers to human and product scent problems because the molecular overlap is high and the universe of relevant molecules is finite.
“There's not an infinite number of molecules out there. There's a lot. But if you get really good at one domain of scent, it turns out to help you in other adjacent domains.”
Osmo's first fully teleported scent was a fresh summer plum, captured from a sliced fruit in a vial, analyzed, and reprinted on the other side of the lab, establishing that a human-perceived smell could be digitized and reconstituted.
“What we did, the first scent that we fully teleported digitally, was a fresh summer plum.”
The analytical instruments for chemistry and scent were already excellent, with about 12 Nobel prizes' worth of advances inside them, so Osmo kept the hardware and replaced the software brains that link the instrument outputs to scent.
“the hardware is already pretty good in the realm of scent and chemistry. But the software, or the maps that link the different pieces of hardware, has been completely missing.”
Osmo's map of scent is about 300-dimensional, unlike the three-dimensional RGB color space, and this higher dimensionality is a reason scent had to wait for AI to be digitized.
“Our map is not three-dimensional like RGB; it's about 300-dimensional, which is, I think, why scent had to wait for artificial intelligence to be digitized”
Fragrance houses are manufacturing businesses with non-manufacturing margins because the blending know-how that produces the exact molecule mix is a deeply guarded secret, and Osmo believes AI can augment that know-how.
“Fundamentally these are manufacturing businesses with non-manufacturing margins because there's a ton of know-how”
Osmo's edge was transferring graph neural network methods, developed for drug discovery by others, into scent, combined with knowing where to find, license, and treat the relevant datasets.
“The intellectual arbitrage that we did was taking those techniques and applying them to the realm of scent.”
Rather than climbing directly toward the full goal of digitizing smell, Osmo finds a route with a shallow enough grade to earn revenue along the way, making the company harder to kill and more likely to survive long term than riskier startups.
“build along a responsible path that makes you harder to kill over time as opposed to making your likelihood of success even riskier over time.”
The long-term vision is a handheld device that reads chemical signals to capture scent, flag health concerns, and recommend vitamins, paired with a device that prints scent, so that pocket-sized sensing becomes ubiquitous.
“The mountain peak is something you can hold in your hand—like your AirPods and your phone—that can read the chemical slice of reality”
Generation is positioned as the near-term commercial vehicle that makes scent design accessible, and it creates the platform and data that will later support the wellness and health frontier.
“that platform is going to help us on our mission to understand human wellness with scent.”
The main existential risk is that nature imposes an unpredictable barrier, such as failing to shrink sensors affordably or a fundamental unknown about how smell works, so Osmo is planning a path of incremental commercial milestones rather than a single leap.
“We could fail to make these sensors small enough to be held in your hand at an appropriate price.”
Companies already use signature scents such as the Ritz-Carlton's, and Alex believes scent is the most emotional brand modality, so more businesses will adopt it as scent becomes accessible.
“But there is no modality, no sense that is more emotional and has deeper ties and associations it can build than scent.”
AromaRama-style theater scent systems in the 1980s failed because they only released a few preset whole-scene smells rather than primary odors that could be mixed to create any scent, which Alex calls a slideshow rather than a display of anything.
“They only had certain smells that were whole scenes. They didn't have primary odors. They didn't have the ability to create any smell.”
Most fragrance houses, after receiving a brief, typically serve the customer a scent from their existing library roughly 90 percent of the time, rather than creating a new custom scent.
“Usually 90% of the time, they go to their library and say, 'Oh, we've already made something for somebody else.”
The current reader must be reduced by a factor of 4 to 8 before it is portable, and the scent printer, currently about half a table in size, also needs substantial miniaturization.
“We have to take that down by a factor of 4 to 8 in order for this to be something you would say is portable right now.”
Scent has strong effects on mood and aromatherapy has a deep tradition, but the underlying science remains underdeveloped, which Osmo believes it can contribute to.
“Scent has powerful effects on our mood, and there's already a very deep tradition of aromatherapy. The science needs to be expanded”
Computers can now see, hear, and touch, but not smell, even though smell feels effortless to humans, and this gap exists because smell's chemical complexity made it extremely hard to digitize.
“But computers can't smell. And that's weird because it's such a fundamental thing.”
A wide book of business with parts that fluctuate with the macro cycle gives a fragrance supplier a resilient profile, and this structure is what Osmo aims to replicate.
“If you build this very wide book of business that has different parts that fluctuate based on the macro, you have a really resilient business there.”
Metropolis, which began with parking-lot management software, became a parking-lot operator when its industry was not ready to buy software, and Alex uses this as precedent for software companies entering a market directly.
“The parking lot industry just wasn't ready as buyers of that software, but it worked. It made parking lots more efficient. So they became a parking lot company.”
Osmo's training labels come from trained human panels rating odors on a rate-all-that-apply scale, and the best label estimate is the average of a panel, which is how high-quality data is produced for AI.
“That's called 'rate all that apply.' It's how we label data.”
About 90% of the products in a household contain a fragrance, and each of those fragrances was crafted by a person combining ingredients, which makes fragrance a very large hidden market.
“So 90% of the products in your home have a fragrance. And every single one of those fragrances was crafted by an individual.”
A perfumer's organ holds about 1,200 ingredient bottles, compared with 88 keys on a piano, and a perfumer combines them in precise ratios to recreate a scent memory.
“Each bottle is like a key on a piano, and instead of 88 keys, it's about 1200 keys.”
Bulgari Black showed Alex that fragrance is an art because it unfolds in three acts over time, which made him realize someone had deliberately designed the sequence.
“this is the fragrance unfolds over time. That was the turning point for me.”
Ilya Sutskever stated at NeurIPS that there is one internet and it has been trained on, so there is no more data for text models, meaning further text gains depend on tricks such as cheaper models and reasoning rather than raw scale.
“Ilya Sutskever... got up at the main AI conference NeurIPS and said, 'Look, there's one Internet and we've trained on it. There's no more data.'”
As authentication-by-smell spreads, counterfeiters will likely start engineering their own scents, creating an arms race that Alex believes is already underway and will be the next frontier.
“Yep, there already is an arms race. This I'm sure is going to be the next frontier”
The fragrance business has operated the same way for about 300 years, and Osmo sees an opportunity to transform it for the better through transparency, safety, and speed, while keeping its time-tested parts.
“The way that business is done here hasn't transformed for 300 years.”
Once Osmo's essential oil delivery matures, candles, detergents, and infusers could carry custom-designed scents that consumers create, making scent design accessible well beyond fine fragrance.
“people will be able to design scents very soon.”
Osmo runs its molecular analysis instruments 24 hours a day with robotic autoloaders and maintains sensory panels twice a day, treating labeling and sensing as scaled, always-on data production.
“Twice a day, we run sensory panels where we just sniff stuff and label stuff.”
Alex believes that as sensors become pocket-sized, computers that can smell will change society by adding information flow and memory, which he argues will largely be a positive development.
“I think it's largely better; there's more information flowing. You're remembering more things.”
Alex intends to run Osmo for his whole life rather than flip it, so he designs the company to survive, citing David Senra's idea that the exit strategy is death.
“the exit strategy is death.”
Osmo's model mirrors drug discovery: AI designs new odorant molecules to customer specifications, such as altering a detergent's smell or replacing a regulated molecule, which are then synthesized and tested.
“Just like a drug company creates a new molecule to affect human health, we create new molecules that affect human perception.”
A computer olfactory system could eventually detect harmful substances at borders and assist or replace sniffer dogs, extending scent detection to truth-in-safety applications.
“I think we could be detecting harmful substances at the border and stopping them. Whatever sniffer dogs are doing, I think eventually a computer will be able to either help with or do it entirely.”
A gas chromatograph olfactometer lets a person pause a blended smell and sniff individual molecules, which Alex compares to debugging software, helping build intuition for new protocols.
“It's like a debugger for software.”
In gas chromatography mass spectrometry, identifying a molecule from its fragment weights and retention time has traditionally required a mix of software and human experts, and Osmo made that interpretation entirely software-driven.
“Typically, this interpretation is done partly by software and partly by people. And what we've done at Osmo is make that happen entirely by software.”
Problems that people think are totally intractable, such as digitizing scent with AI, become tractable once one starts working on them, because progress compounds bit by bit.
“problems that people sometimes think are totally intractable. Once you get started, you realize we're actually making progress.”
Once sensors fit in a pocket, people may hold onto moments more richly through captured scents, creating beauty and more information flow, although the effects on society are hard to predict.
“You're remembering more things. You can hold on to moments—there's some beauty that's there that wasn't there before”
Osmo's founding path ran through Google Brain, where Alex and colleagues spun out the project after Josh Wolfe of Lux Capital heard a pitch on a decade-long thesis about digitizing olfaction, leading to a funded company with Lux leading and GV co-leading.
“Josh led the round, Krishna at GV co-led, and we built Osmo and we're on our way.”
People often sense when a partner is getting sick before symptoms appear, which Alex takes as evidence that the body's chemical state is detectable by smell.
“a lot of people can sometimes smell if their loved one or their partner is getting sick or if something's a little bit off and then two days later they actually get sick. That's real.”
Osmo is an AI company in the strict sense, with an org chart that looks like an AI company married to a chemistry company, unlike most AI application companies that remix existing models and software.
“One of the things that's so interesting to me about Osmo is it is an AI company in a very strict sense.”
The fidelity gap between read and write is measured by whether a familiar smell matches a remembered one, using debiasing tricks, with the human customer as the final arbiter.
“But you're the arbiter. If it's a smell you're familiar with, and I'm trying to recreate a memory that you have, then we either did it or we didn't.”
Osmo has proven that product scent can reveal provenance, and this is already deployed at StockX, which is a validated commercial result rather than a lab concept.
“We've proved out that we can use the scent of a product to tell whether it's a real or a fake”
Osmo can analyze any object's smell without liquid extraction by pumping air into a vial, concentrating the odor onto absorbent film, and thermally desorbing it into the spectrometer, enabling analysis of flowers, vegetables, people, and fruit.
“This means that we can analyze flowers, vegetables, people, and fruits.”
Osmo uses several kinds of AI models and LLMs where useful, but it avoids forcing every problem into a language model, following a 'right tool for the job' philosophy.
“Our philosophy is the right tool for the job. And so we're not going to take a dogmatic approach and try to shove everything into an LLM.”
Alex's interest in perfume began at age 12 when he noticed fragrance changed how people treated him, and a Bulgari Black spray unfolding over time convinced him perfumery was an art form, leading him toward neuroscience.
“I thought, 'No, someone made this.' This is the whole thing—the fragrance unfolds over time.”
Generation is Osmo's fragrance house that combines its scent-digitization technology with human perfumers and noses to let creators and brands design custom fragrances quickly.
“The idea is to take all the technology that we've built, but also all the humanity—people, the perfumers, the noses”
Nearly every line of Osmo's code is touched by AI tools, which raises productivity and means Osmo is a smaller company than it otherwise would be.
“I don't think there's any code that we write that isn't at least touched by AI in some way. Osmo's probably a smaller company as a result of all these AI tools.”
Alex credits the kindest thing anyone has done for him as the willingness of mentors and backers, from academia through Google to Osmo's board, to bet on him when evidence was thin.
“it's these little moments where people say yes that can make all the difference.”
Alex started and sold two AI companies, one in biotech and one in ML-as-a-service, with the ML company bought by Twitter, where he helped build a deep learning team that applied AI to ad platforms and data centers.
“I started and sold two AI companies—one in the biotech space and one in the pure ML-as-a-service space. That ML company was bought by Twitter.”